Speech analysis and synthesis using a vocal tract/cord model

Abstract

The use of our vocal cord and tract model [Schroeter and Sondhi, J. Acoust. Soc. Am. Suppl. 1 78, S6 (1985)] for speech coding at bit rates below 4.8 kb/s is being considered. Earlier work [Flanagan et al., J. Acoust. Soc. Am. 68, 780–791 (1980)] reported on applying an unconstrained optimization method for jointly finding the best vocal tract and glottal parameters directly from the speech signal (one‐stage optimization). This method has been extended by optimizing tract and glottal parameters sequentially (two‐stage optimization). In a first stage, a search was made for the parameters of an inverse (tract) filter that give the best match between the inverse filtered speech and the output of the vocal cord model for a given set of glottal parameters. In a second stage, the parameters of the cord model were optimized for a given set of tract parameters so that the synthesized speech matches the original speech as closely as possible. This method has the advantage of separating the source and tract into different optimization stages. By doing so, the inherent constraints of the respective models can be easily used to constrain the search space. Compared to the one‐stage optimization, the search required in each stage is over a parameter space of a smaller dimension. Different start‐up procedures for the two‐stage optimization are investigated. Results are compared to those of the one‐stage optimization.